Hmm, doesn't look like it's shaping up to be a card for the enthusiasts.
Once combined cpu/larrabee chips it'll hopefully spell the end for completely useless gfx on netbooks - after all intels pricing policy makes ion pretty prohibitive!!
Quote:
Originally Posted by article we're not getting our hopes up about Larrabee's 3D grapihcs performance
"I still don't get Intel's obsession with ray tracing"
Because ray tracing or similar techniques are the "right way" to do things. Raster graphics are a huge collection of hacks. To my knowledge, any decent solution for radiosity, reflection and refraction with raster is either a huge hack or starts approaching tracing techniques. You get all that for free with ray tracing.
Also, ray tracing requires no special-purpose hardware. It's extremely easily to parallelize. Intel are the guys that want to make 80 core general purpose CPUs, remember? Massively parallel, general purpose is within Intel's grasp (5-10 years?) so it makes sense.
For what it's worth I'm somewhat impressed considering it's realtime raytracing, which is a fundamentally different (and computationally intensive) approach to 3D rendering than the rasterization approach we're used to. Have to admit that I'd like to see how Larrabee would perform using that approach.
Originally Posted by Emon "I still don't get Intel's obsession with ray tracing"
Because ray tracing or similar techniques are the "right way" to do things. Raster graphics are a huge collection of hacks. To my knowledge, any decent solution for radiosity, reflection and refraction with raster is either a huge hack or starts approaching tracing techniques. You get all that for free with ray tracing.
It might be "the right way" to do things, but please tell me a game shipping today which uses ray tracing. For a product shipping in the middle of 2010, ray tracing is a pointless demo... in fact I'd go so far as say it's "the wrong way" to show Larrabee off. Please show me, and other fellow gamers, how it plays the games I play today.
Secondly, when it runs as slowly as it did in the demo, it does nothing but prove that a chip supposedly designed for ray tracing can't even render a basic scene (that looks dated) in "the right way"... again, that smells of "the wrong way" to me again. Raster graphics may be a huge collection of hacks, but it is what games use today - the benefit it has is that performance is uncompromised and that's why developers use it. Performance is hugely compromised with a completely ray traced environment, as proven by Intel's demo - the water looks much more dated and so do the ripples. Not even film animators use ray tracing exclusively and they measure performance in frames per day, not frames per second.
I'm not against ray tracing, but I don't see the obsession with ray tracing every part of a scene - it's hugely expensive and there is, frankly, no benefit. There are benefits to ray tracing curved surfaces, transparent surfaces and much more, but compute power needs to increase by a factor of thousands before ray tracing is viable in a real-time environment.
For what it's worth I'm somewhat impressed considering it's realtime raytracing, which is a fundamentally different (and computationally intensive) approach to 3D rendering than the rasterization approach we're used to. Have to admit that I'd like to see how Larrabee would perform using that approach.
Intel has been demoing real-time ray tracing for years at higher and higher resolutions, running on the CPU only in the past. The fact that a 6-core, 12-thread Gulftown, PLUS Larrabee, ran the demo worse than an 8-core, 16-thread box with ATI/Nvidia graphics last year (60fps at 720p, fwiw) is very, very disappointing. Of course, Intel won't disclose the resolution it was running at - I'd guess 1080p - but it wasn't even running at anything close to 30fps and with a completely static camera unlike previous demos.
So are you saying that Larrabee is running slower that the 8 core 16 thread processor of last year, or was the ray tracing running on the graphics card? Either way that really sucks.
Originally Posted by Cupboard So are you saying that Larrabee is running slower that the 8 core 16 thread processor of last year, or was the ray tracing running on the graphics card? Either way that really sucks.
I believe the demo is running at a higher resolution, but I don't know what resolution it is running at (Intel won't say), but it wasn't running at 30fps, never mind the 60fps that we saw last year. I'm guessing it was 1080p based on the size of the screen relative to other stuff around it, which is effectively 2.25x the resolution of 720p (0.92mp vs 2.07mp)... core count (on the CPU) reduced by 25 per cent, but they've added Larrabee on top... which has many more 'cores' - it would have been an impressive ray tracing demo if it hit 60fps at 1080p with early Larrabee silicon, but it didn't. The fact that it didn't with a static camera just compounded that disappointment.
As a result, you've not only got the let-down of not finding out how Larrabee is going handle Direct3D applications (the big question mark for most of us, I'm sure... even if it's just smooth vs. not smooth), but you've also got the compounded disappointment of the demo looking dated, being a static camera angle and not running particularly well (which isn't a great proof of concept for real-time ray tracing being 'just around the corner'). The water seriously doesn't look that much better than the water in Far Cry - a game released on DX9 in 2004. :)
For what it's worth I'm somewhat impressed considering it's realtime raytracing, which is a fundamentally different (and computationally intensive) approach to 3D rendering than the rasterization approach we're used to. Have to admit that I'd like to see how Larrabee would perform using that approach.
To me, your comments on ray-tracing are spot on, Tim. It may in a sense become more efficient than rasterisation at higher resolutions, but when will we be playing games at such resolutions and have the hardware to do so? No time soon, for certain. When ray-tracing is introduced for games, it will be in tandem with rasterisation without doubt.
I really don't see how Larrabee can succeed. With respect to games, I know that Tim Sweeney has said that he sees games being rendered through software in the future but even then, Larrabee would surely not be the hardware to do it. If he was to commit himself to designing a games engine along these lines at this stage, I think it would be a waste. I don't see why he would anyway because surely it would not be able to be used for consoles (this or the next generation), even if there are reports that the Playstation 4 will use Larrabee for its graphics.
With respect to general computational work, surely GPUs are more efficient in terms of speed and power anyway, whether or not they are harder to code for.
No matter what it is used for, Larrabee will be too big and hot to be worthwhile, or so it would seem.
(Hope you're feeling better too).
Intel appears to be threading the path of disruptive innovation with Larrabee - a path where lower performance is expected at the early stages of the technology but - hopefully - incremental development will allow the new tech to catch up with the average needs of the GPU market.
However, if initial performance is not even capable of appealing to the lower end of the market, and no developers adopt ray tracing (and I aggre with Tim - it is not about the "right" way to do things, but rather the appropriate way to do things, the way that gets the job done), I cannot see much future in Larrabee.
Intel is certainly not dumb. Larrabee is a huge gamble that may pay off in the long run. But "underwhelming" is certainly not the word we were all expecting after an year and a half of hype around this new architecture.
I see many ppl are discussing about the ray-tracing.
yes, it INDEED is one fun point of Larrabee, but not all.
It's just a derivation of Larrabee's software rendering revolution.
Larrabee will be the first gfx card that's built upon x86 architect. The vast differences between Larrabee and traditional GPUs are programbility.
Larrabee enables us to code the whole render pipeline almost entirely in software, thus dramatically increases the complexity of output images while maintain reasonable performance.
Sadly, intel tries its best to cover all tech details.
So, its actual performance on next-gen games is unknown.
Talking about ray-tracing, surly itl be slower than rasterization method if only local direct lighting is taken into concern.
wow im surprised. i was thinking larrabee would be demoing something like the demo ASUS had for the Mars. it seems like intel is putting all their eggs into the ray tracing basket. kinda risky move
Ok so the demo was a bit poor, however it is nice to see they are trying the new thing. They may have accepted they can not compete easily in the D3D arena with nVidia and AMD not really trying anything new, other than making the chips faster and with more memory, the cards are still essentially doing the same thing they have done for years.
Raytracing will happen at some point, and I can see when Intel crack it and if they release good tools for it, then nVidia/AMD might be left stuck, however since Labree is a paraelle processing card, the theroy would be once Intel have it all working, they should be able to replease an API that Cuda or whatever the AMD version is called, to run ray tracing.
Larabees a bit of a disspointment in terms of its ability solidly chuck out the fps, i was expecting more, especially with some of the hype on the internet, and intels own bragging.
But at least its a step forward for ray tracing, if not a pointless one for the consumer market
Larrabee is not really about producing a graphics card. Its about producing a highly parallel set of CPUs with good floating point throughput that is x86 compatible. The graphics card vendors have seen fit to start to target the high performance market, traditionally Intel's market and this is Intel's response to CUDA, OpenCL and DirectCompute running on ATI/NVidia graphics cards. I don't think the end goal is to run a full directX game on this architecture yet, just to secure the HC market against the intrusions of the less capable graphics cards. This is a market that Intel wants to dominate as its only now emerging.
One such program running on that architecture is a graphics rendering engine and ray tracing is a much simpler engine to implement to make it look good than a fully raster capable one. I expect we'll see other technical demos from Intel as we approach the launch that are all about high computation and massively parallel problems, with a focus on an API much easier to use than OpenCL and ilk.
Ray tracing and graphics rendering is in many ways a warning to the graphics card vendors, especially NVidia. The war of words 2 years ago was that NVidia was going to eat Intel with its massively parallel stream processors, now Intel is not only producing a significantly faster and more capable multi core system but its also showing how one day it'll be able to eat NVidia's market.
So in a nutshell don't expect Larrabee to really beat out the 5870, its likely to be an order of magnitude slower in game rendering. However in the high compute market Larrabee looks like its going to dominate. Moore's law will then (Intel hopes) take it to eating the NVidia/ATI graphics market with engines that are fully programmable and with higher fidelity with Ray Tracing, all implemented in Intel's key IP - x86.
Originally Posted by BrightCandle Larrabee is not really about producing a graphics card.
Every piece of Intel material would disagree with that statement. The first products are "discrete graphics" and support DirectX, OpenGL and OpenCL (as well as the native programming mode) which, if I'm not mistaken, means it's a graphics card. I wish people would stop trying to say it's something else. :(
Time is on Intel's side, that's the reason for "Intel's obsession with ray tracing". As cores and threads increase with each new generation of CPUs, ray tracing will approach viability and, eventually, even surpass the DirectX version of the day. It's a problem to be solved by adding muscle. Yes, right now it's nothing but marketing hype: Larrabee won't even come close to compete with what ATI and Nvidia had last year, much less with future GPUs. But Intel have deep pockets and Larrabee is but the first step. A very important first step.
I think tim has a point that kind of stuck with me least.. ray tracing is wasteful- might be easier on the programmer but it takes alot more power to render the same scene
I'm not sure what Intel is trying to do with Larrabee. It almost looks like Intel doesn't know what Intel is doing with it either.
Raytracing is amazingly difficult and computationally expensive to do right, but it seems like the demo was solely to 'justify' Larrabee existance at the moment. I agree that the pseudoGPU actually has relevance, but where it will fit in eventually is a bit of a crap shoot.
At 32 cores, Larrabee hits about 2 SP TFLOPs @ 2GHz whereas the Radeon 5870 sits in at 2.72TFLOPs. If Intel is going down the "graphics card" road with this, they'll have one heluva problem...mainly, their GPU doesn't compete at the moment.
Of course, to hit the same theoretical bandwidth, you're looking at a 48-core Larrabee at 2GHz* using quite a bit more power (assuming both hit maximum theoretical performance on both cards and being generous on the power specs for Larrabee).
* - Assuming that Larrabee scales at 100% to 32 cores, and 90% after that.
Comments 1 to 25 of 30
ReplyOnce combined cpu/larrabee chips it'll hopefully spell the end for completely useless gfx on netbooks - after all intels pricing policy makes ion pretty prohibitive!!
this reminds me alot of the volari line of cards that came out few years back.. they were so weak what was the point
i suppose intel cant win every time, that demo must have really sucked!
peace
f.
Because ray tracing or similar techniques are the "right way" to do things. Raster graphics are a huge collection of hacks. To my knowledge, any decent solution for radiosity, reflection and refraction with raster is either a huge hack or starts approaching tracing techniques. You get all that for free with ray tracing.
Also, ray tracing requires no special-purpose hardware. It's extremely easily to parallelize. Intel are the guys that want to make 80 core general purpose CPUs, remember? Massively parallel, general purpose is within Intel's grasp (5-10 years?) so it makes sense.
Very disappointing indeed :-(
I was hoping Intel would WOW everyone.
Maybe, just maybe things will change right at the end - hehe
For what it's worth I'm somewhat impressed considering it's realtime raytracing, which is a fundamentally different (and computationally intensive) approach to 3D rendering than the rasterization approach we're used to. Have to admit that I'd like to see how Larrabee would perform using that approach.
It might be "the right way" to do things, but please tell me a game shipping today which uses ray tracing. For a product shipping in the middle of 2010, ray tracing is a pointless demo... in fact I'd go so far as say it's "the wrong way" to show Larrabee off. Please show me, and other fellow gamers, how it plays the games I play today.
Secondly, when it runs as slowly as it did in the demo, it does nothing but prove that a chip supposedly designed for ray tracing can't even render a basic scene (that looks dated) in "the right way"... again, that smells of "the wrong way" to me again. Raster graphics may be a huge collection of hacks, but it is what games use today - the benefit it has is that performance is uncompromised and that's why developers use it. Performance is hugely compromised with a completely ray traced environment, as proven by Intel's demo - the water looks much more dated and so do the ripples. Not even film animators use ray tracing exclusively and they measure performance in frames per day, not frames per second.
I'm not against ray tracing, but I don't see the obsession with ray tracing every part of a scene - it's hugely expensive and there is, frankly, no benefit. There are benefits to ray tracing curved surfaces, transparent surfaces and much more, but compute power needs to increase by a factor of thousands before ray tracing is viable in a real-time environment.
Intel has been demoing real-time ray tracing for years at higher and higher resolutions, running on the CPU only in the past. The fact that a 6-core, 12-thread Gulftown, PLUS Larrabee, ran the demo worse than an 8-core, 16-thread box with ATI/Nvidia graphics last year (60fps at 720p, fwiw) is very, very disappointing. Of course, Intel won't disclose the resolution it was running at - I'd guess 1080p - but it wasn't even running at anything close to 30fps and with a completely static camera unlike previous demos.
I believe the demo is running at a higher resolution, but I don't know what resolution it is running at (Intel won't say), but it wasn't running at 30fps, never mind the 60fps that we saw last year. I'm guessing it was 1080p based on the size of the screen relative to other stuff around it, which is effectively 2.25x the resolution of 720p (0.92mp vs 2.07mp)... core count (on the CPU) reduced by 25 per cent, but they've added Larrabee on top... which has many more 'cores' - it would have been an impressive ray tracing demo if it hit 60fps at 1080p with early Larrabee silicon, but it didn't. The fact that it didn't with a static camera just compounded that disappointment.
As a result, you've not only got the let-down of not finding out how Larrabee is going handle Direct3D applications (the big question mark for most of us, I'm sure... even if it's just smooth vs. not smooth), but you've also got the compounded disappointment of the demo looking dated, being a static camera angle and not running particularly well (which isn't a great proof of concept for real-time ray tracing being 'just around the corner'). The water seriously doesn't look that much better than the water in Far Cry - a game released on DX9 in 2004. :)
Thanks for the link !
I really don't see how Larrabee can succeed. With respect to games, I know that Tim Sweeney has said that he sees games being rendered through software in the future but even then, Larrabee would surely not be the hardware to do it. If he was to commit himself to designing a games engine along these lines at this stage, I think it would be a waste. I don't see why he would anyway because surely it would not be able to be used for consoles (this or the next generation), even if there are reports that the Playstation 4 will use Larrabee for its graphics.
With respect to general computational work, surely GPUs are more efficient in terms of speed and power anyway, whether or not they are harder to code for.
No matter what it is used for, Larrabee will be too big and hot to be worthwhile, or so it would seem.
(Hope you're feeling better too).
However, if initial performance is not even capable of appealing to the lower end of the market, and no developers adopt ray tracing (and I aggre with Tim - it is not about the "right" way to do things, but rather the appropriate way to do things, the way that gets the job done), I cannot see much future in Larrabee.
Intel is certainly not dumb. Larrabee is a huge gamble that may pay off in the long run. But "underwhelming" is certainly not the word we were all expecting after an year and a half of hype around this new architecture.
yes, it INDEED is one fun point of Larrabee, but not all.
It's just a derivation of Larrabee's software rendering revolution.
Larrabee will be the first gfx card that's built upon x86 architect. The vast differences between Larrabee and traditional GPUs are programbility.
Larrabee enables us to code the whole render pipeline almost entirely in software, thus dramatically increases the complexity of output images while maintain reasonable performance.
Sadly, intel tries its best to cover all tech details.
So, its actual performance on next-gen games is unknown.
Talking about ray-tracing, surly itl be slower than rasterization method if only local direct lighting is taken into concern.
Dex
Raytracing will happen at some point, and I can see when Intel crack it and if they release good tools for it, then nVidia/AMD might be left stuck, however since Labree is a paraelle processing card, the theroy would be once Intel have it all working, they should be able to replease an API that Cuda or whatever the AMD version is called, to run ray tracing.
But at least its a step forward for ray tracing, if not a pointless one for the consumer market
One such program running on that architecture is a graphics rendering engine and ray tracing is a much simpler engine to implement to make it look good than a fully raster capable one. I expect we'll see other technical demos from Intel as we approach the launch that are all about high computation and massively parallel problems, with a focus on an API much easier to use than OpenCL and ilk.
Ray tracing and graphics rendering is in many ways a warning to the graphics card vendors, especially NVidia. The war of words 2 years ago was that NVidia was going to eat Intel with its massively parallel stream processors, now Intel is not only producing a significantly faster and more capable multi core system but its also showing how one day it'll be able to eat NVidia's market.
So in a nutshell don't expect Larrabee to really beat out the 5870, its likely to be an order of magnitude slower in game rendering. However in the high compute market Larrabee looks like its going to dominate. Moore's law will then (Intel hopes) take it to eating the NVidia/ATI graphics market with engines that are fully programmable and with higher fidelity with Ray Tracing, all implemented in Intel's key IP - x86.
Every piece of Intel material would disagree with that statement. The first products are "discrete graphics" and support DirectX, OpenGL and OpenCL (as well as the native programming mode) which, if I'm not mistaken, means it's a graphics card. I wish people would stop trying to say it's something else. :(
Raytracing is amazingly difficult and computationally expensive to do right, but it seems like the demo was solely to 'justify' Larrabee existance at the moment. I agree that the pseudoGPU actually has relevance, but where it will fit in eventually is a bit of a crap shoot.
At 32 cores, Larrabee hits about 2 SP TFLOPs @ 2GHz whereas the Radeon 5870 sits in at 2.72TFLOPs. If Intel is going down the "graphics card" road with this, they'll have one heluva problem...mainly, their GPU doesn't compete at the moment.
Of course, to hit the same theoretical bandwidth, you're looking at a 48-core Larrabee at 2GHz* using quite a bit more power (assuming both hit maximum theoretical performance on both cards and being generous on the power specs for Larrabee).
* - Assuming that Larrabee scales at 100% to 32 cores, and 90% after that.
-
« Previous
-
1
-
2
-
Next »
Discuss in the forums